View
9
Download
0
Category
Preview:
Citation preview
Deterring or Displacing Electoral Irregularities?
Spillover Effects of Observers in a Randomized Field
Experiment in Ghana
Nahomi Ichino∗ Matthias Schundeln†
15 July 2011
∗Assistant Professor, Department of Government, Harvard University, 1737 CambridgeSt., Cambridge, MA, 02138. E-mail: nichino@gov.harvard.edu.†Professor, Faculty of Economics and Business Administration, Goethe University Frank-
furt, RuW, Postbox 46, Gruneburgplatz 1, D-60323 Frankfurt am Main, Germany. E-mail:schuendeln@wiwi.uni-frankfurt.de.
1
Abstract
This article studies the effect of domestic observers deployed to reduce irregularities in
voter registration in a new democracy, and in particular, the response of political parties’
agents to these observers. Because political parties operate over large areas and party agents
may relocate away from observed registration centers, observers may displace rather than
deter irregularities. We design and implement a large-scale two-level randomized field exper-
iment in Ghana in 2008 taking into account these spillovers and find evidence for substantial
irregularities: the registration increase is smaller in constituencies with observers; within
these constituencies with observers, the increase is about one-sixth smaller on average in
electoral areas with observers than in those without; but some of the deterred registrations
appear to be displaced to nearby electoral areas. The finding of positive spillovers has im-
plications for the measurement of electoral irregularities or analysis of data collected by
observers.
Keywords: democratization, voter registration, field experiment, spillovers, electoral fraud
2
Following the third and fourth waves of democratization of the late twentieth and early
twenty-first century, an overwhelming number of countries in the world today elects its
leaders. In these new democracies, popular elections are frequently marked by fraud and
irregularities (Simpser 2010), which affect public confidence in democracy and regime legiti-
macy (Elklit and Reynolds 2002; Birch 2008; Rose and Mischler 2009), political participation
(McCann and Domınguez 1998), and protest and political violence (Eisenstadt 2002). For
both historical and contemporary cases, the question of how informal practices of ballot
stuffing, registration fraud, and other electoral malpractices are eliminated are now central
to the study of democratization (Ziblatt 2006), which had earlier focused on changes to
formal rules like the extension of the suffrage or the development of responsible and limited
government. An emerging body of scholarship on democratization and new democracies
argues that the extent of electoral fraud is affected by political competition (Lehoucq 2003;
Simpser 2010), electoral rules (Birch 2007; Lehoucq and Molina 2002), socio-economic in-
equality (Ziblatt 2009), the quality of the electoral management body that organizes and
conducts the elections (Elklit 1999; Elklit and Reynolds 2002; Hartlyn et al. 2008; Pastor
1999), and scrutiny by international election monitors (Hyde 2007; Kelley forthcoming).
The fundamental difficulty with the study of election fraud is its measurement – it may
take many forms and those involved typically wish to hide these illicit activities. Scholars
generally rely upon assessments by election observers to measure electoral fraud for quanti-
tative cross-national studies and use media reports and petitions filed by aggrieved parties
for single-country studies (Lehoucq 2003). But these measures are generated by participants
with different interests, expectations and standards across elections (Kelley forthcoming),
which raises concerns about consistency and bias. Moreover, politicians may respond to the
possibility of detection by observers or media by engaging in fraud at alternative locations
or earlier stages of the electoral process that are less likely to be detected (Bjornlund 2004;
Carothers 1997), much as police and other crime-fighting measures have spillover and dis-
3
placement effects on criminal activity (Bronars and Lott 1998; Di Tella and Schargrodsky
2004). Consequently, current measures used in cross-national studies may underestimate the
extent of electoral fraud in new democracies, including those that appear to have fairly clean
elections. This underlies the need for substantial caution about the robustness of empirical
findings in this emerging large-n literature (Birch 2007).
This article studies empirically the strategic response of political parties to civil society
actors’ efforts to detect and deter misconduct ahead of a closely contested election in a new
democracy. More specifically, we design and implement a randomized field experiment to
examine the causal effect of domestic election observers on the extent of irregularities in voter
registration conducted over a 13-day registration period in advance of the 2008 Ghanaian
general elections. We directly address the aforementioned measurement problem and the
violation of the stable unit treatment value assumption (SUTVA, Rubin 1978) implied by
the possible displacement of irregularities using a two-level randomization design. This
design enables us to detect, and indeed we find, localized and general spillover effects that
are consistent with evasive responses by the political parties to the observation effort of
the Coalition of Domestic Election Observers (CODEO) in Ghana. Rather than deterring
irregularities entirely, observers displace a substantial portion of irregularities to nearby
unobserved registration centers in a pattern consistent with communication among political
party agents. The effects of these observers and the estimated extent of irregularities are
substantial, buttressing Birch’s (2007) concerns.1
We focus on voter registration for two reasons. First, problems with voter registration are
quite common in transition elections and new democracies, and this can create significant
doubts about electoral outcome and the legitimacy of the new government. The 1991 general
elections in Nepal, the first free elections in that country in over 30 years, were held using a
register in which the number of registered voters exceeded census estimates by about 10–15%
(Gaige and Scholz 1991, 1056). Similarly, the Philippine general elections in 1998 were run
4
by an electoral commission that refused to “reorganize old voter lists or issue identification
cards, thus leaving the door open for so-called “flying voters” (i.e., those who vote more
than once)” (Case 1999, 474). Substantial problems with the voter registration process and
the resulting voters register also marred the 1993 elections in Senegal. There were numerous
claims of discrimination against opposition supporters in the provision of voter cards and an
estimated 30–50% of voter cards had factual errors which could prevent people from voting.
Moreover, “public perception” that the incumbent government had abused the system of
documents that allowed people whose names did not appear on the register to vote “was
undoubtedly the single most harmful issue in terms of eroding public confidence in the
integrity of the elections” (Villalon 1994, 178).
In more serious cases, voter registration problems may cause elections to be delayed and
possibly not held at all. Accusations of fraud and violence around voter registration in
2008, for example, eventually led to the postponement of parliamentary elections in Yemen
from 2009 to 2011 (AFP 2009; Yemen Times 2009). Similarly, the election crisis in Cote
d’Ivoire in late 2010 was presaged by the inability of the government and opposition to
agree to procedures for voter registration. This led to the dismissal of both the Electoral
Commission and the government, a delay of the highly anticipated first election since the
end of the civil war, and deadly demonstrations (RFI 2010; Zamble 2010).
Second, political parties have strong incentives to inflate the voters register with their
own supporters, even where they are unable to fabricate elections results outright or to
widely intimidate voters and opponents, which is precisely where the election would likely
be characterized as “free and fair.” Political parties can skew the results in their favor on
election day by enabling multiple-voting or adding pre-marked ballot papers to ballot boxes,
without obviously pressuring a voter or restricting his access to a polling station. The risk
to this strategy, however, is that if the number of votes cast as a proportion of registered
voters appears suspiciously high, public scrutiny and legal and political challenges will follow.
5
To benefit from the extra votes then, a party must ensure that extra names appear on the
voters register. Consequently, incentives to rig elections become incentives to inflate the
voters register, and moreover, to evade the scrutiny of observers while doing so. Organized
political parties should then have party agents avoid the registration centers where observers
are located and instead try to inflate the register at nearby unobserved registration centers.
Our study is sited in Ghana in sub-Saharan Africa because it shares with many other
partial democracies political and institutional characteristics that are expected to affect
politicians’ incentives to engage in electoral malpractices (Birch 2007; Hartlyn et al. 2008).2
It has a majoritarian electoral system and an electoral commission that is officially indepen-
dent but under-resourced. Like many partial democracies, Ghana is a rapidly urbanizing
developing country with a large, poor, rural population, where resources are concentrated
in the state – conditions that are frequently associated with vote buying and other electoral
malpractices (Kitschelt and Wilkinson 2007; Stokes 2005).
Furthermore, Ghana has significant experience with both international and domestic elec-
tion observers who, as in other countries, have historically focused on election day activities
and may have pushed malpractices by political parties to the pre-election stage. Indeed, there
is prima facie evidence of inflation of the voters register. In 2008, the Electoral Commis-
sion expected to register 800,000 voters, the estimated number of citizens who had attained
voting age since the previous registration. However, the Electoral Commission registered
nearly 2 million new voters, a figure significantly greater than the vote margin in the previ-
ous presidential election and for a provisional total number of registered voters greater than
the estimated adult population of Ghana. Some but not all of these unexpected 1.2 million
registrations were people mistakenly re-registering instead of requesting a replacement for a
lost voter ID card. However, in several constituencies, the two main political parties traded
accusations of transporting supporters to have them illegally register to vote (Boateng 2008).
6
Our contribution is threefold. First and most immediately, we shift the focus in this
literature on electoral misconduct to the pre-election stage. We demonstrate and quantify
sizable irregularities in voter registration in a country that is considered a “model” new
democracy 16 years removed from its transition from autocratic rule. To our knowledge,
our work is the first large-scale experimental study to examine pre-election irregularities
and to work with domestic election observers. It complements related work by Hyde (2007)
with election-day international monitors and the extensive qualitative reporting on election
fraud by domestic and international election observers supported by organizations such as
the Carter Center, National Democratic Institute, and the European Union.3 Our work
also adds to the literature on statistical methods to detect problems with election results
(Mebane 2006; Myagkov et al. 2009).
Second, we consider explicitly how political parties are organizations that cover a wide
geographical area and create connections across political units, and we study the implication
that interventions on illicit political activities will have spillover effects. We find evidence
for such interference across spatial units, but also conclude that the spillovers are limited by
geographical distance. Our work adds to the handful of recent studies in political science
that have explicitly considered spillovers, such as Nickerson’s (2008) experimental study of
peer effects on turnout in an American election.
Third, this research contributes to a growing body of scholarship that uses randomized
field experiments to study how democracy works in practice in specific developing coun-
tries and clarify debates and generalizations from observational studies (Wantchekon 2003;
Humphreys et al. 2006; Olken 2010; Collier and Vicente 2010). Our experimental study
reinforces Birch (2007)’s concerns about the robustness of findings from large-n empirical
studies that measure fraud using observer reports. These include most obviously stud-
ies on the causes and consequences of electoral fraud, but also those on the relationship
between election quality and democratic development (Lindberg 2006). Our work also sug-
7
gests greater circumspection about the efficacy of observers in reducing electoral irregularities
(Hyde 2007; Kelley forthcoming) and points to future research on how fraud is organized
and on the relationship between pre-election fraud, election day problems, and post-election
rigging.
We proceed by first presenting our hypotheses on observers, party agents, and voter
registration. We describe the voter registration process and the 2008 general elections in
Ghana, then present the experimental design, data, and analysis.
Observers and Party Agents
What is the effect of domestic observers on the behavior of political party agents and the
extent of irregularities in voter registration? The basic premise is that political parties want
to win the election and want to appear to do so cleanly. If the parties intend to inflate the vote
total in their favor on election day, they may wish to inflate the register with their supporters
and at the same time do not want their agents to be caught doing so. Therefore, observers
should, in fact, deter party agents from organizing logistics for fraudulent registrations at
the registration centers that the observers visit. We call this causal effect of observers on
registration at the registration centers to which they are deployed the primary effect. This
logic implies the following hypothesis: A registration center with an observer should have a
lower increase in registration than registration centers without observers (negative primary
effect).
However, two additional effects are possible. Political parties are concerned with voting
and voter registration over a wide area composed of many registration centers and political
party agents often transport their supporters in vans and buses. Party agents can often
communicate with one another and travel fairly easily to avoid a particular registration cen-
ter with an observer. However, time and resource constraints imply that not all deterred
registrants would be relocated to alternative registration centers and whether there is a regis-
8
tration center nearby would affect the extent of this relocation. Consequently, some portion
of the extra registration deterred by an observer may simply be displaced to nearby regis-
tration centers, which we call a localized spillover effect. This implies a second hypothesis:
A registration center with a nearby observed registration center should have a larger increase
in registration than registration centers far away from observed registration centers (positive
localized spillover effect).
The second possible effect is that observers may deter extra registrations in the con-
stituency overall, not just at the registration centers to which they are deployed. The
presence of these registration observers may become widely known and give the impression
that the constituency overall is being observed, so that observers may also have a negative
general spillover effect.
We interpret the primary effect as a lower bound on the extent of irregularities in voter
registration enabled by political party agents, since observers likely do not deter all problems
at the locations they visit. However, observers might affect registration through alternative
mechanisms that would complicate this interpretation. First, citizens who know that an
observer is present at their local registration center may feel less intimidated by possible
trouble and become more likely to register to vote. This would attenuate the primary
effect of an observer on the visited registration center and, consequently, also implies that
a registration center with a nearby observed registration center may not experience a lower
additional increase in voter registration. Second, electoral officials might also become either
more efficient in carrying out their duties or more diligent and slow down the registration
process. If the effect of observers comes through the influence on electoral officials, the
expected primary effect of observers is unclear, and as in the case of influence on citizens, a
registration center with a nearby observed registration center should have a lower additional
increase in voter registration. However, electoral officials who see registration observers at a
registration center could report this up the chain of command, affecting the behavior of their
9
counterparts at other registration centers, so that there may be a general spillover effect.
We check these alternative mechanisms following our main analysis.
Voter Registration and the National Election in Ghana 2008
Ghana is a former British colony and is an ethnically and religiously diverse country of
approximately 23 million on the Gulf of Guinea in West Africa, bordered by Ivory Coast,
Togo, and Burkina Faso. The Akans, who are concentrated in the more prosperous southern
parts of the country, are the largest ethnic group in Ghana, but they are composed of many
distinct sub-groups that together do not comprise a majority of the population.
Ghana has a history of cycles of coups d’etat and military rule, but it has held regular,
competitive elections every four years since its transition to democracy in 1992. Direct
elections are held concurrently for president and a unicameral national parliament, which is
composed of 230 members elected by plurality from single-member districts. The winning
candidate for the presidency must win a majority of votes cast, with a run-off election
between the top two vote-getters if no candidate wins a majority in the first round. The
then incumbent military ruler, Flt. Lt. Jerry Rawlings, was elected the first president of
the Fourth Republic in 1992 and then re-elected in 1996 on the platform of the National
Democratic Convention (NDC). Rawlings left office in 2000 after the constitutional limit of
two terms, and his party’s presidential candidate was defeated by John Kufuor of the New
Patriotic Party (NPP) in a very close election. Kufuor was re-elected in 2004 and there was
no question that he would leave office in 2008 following his two terms. Since 1992, the NDC
and NPP have been the two major parties in Ghana and both parties run candidates and
compete in local-level and national-level elections throughout the country. The NDC and
NPP consider themselves left-leaning and right-leaning, respectively, and each is strongly
identified with regional and ethnic bases (Lindberg and Morrison 2005, Nugent 2001).
10
In order to vote in Ghana, a voter must go to the particular polling station associated
with his residence and present his voter ID card which should have a photograph taken at
the time of registration if a camera was available; no other form of identification is required.
The electoral official compares this card with the information and photograph printed on
the voters register before allowing the person to vote. While international borders are closed
around election day to prevent Togolese, Burkinabe, and other foreigners from entering the
country to vote, internal roads are left open and someone who wishes to vote in multiple
locations on election day could easily do so as long as he is registered at those multiple
locations.
Citizens of Ghana may register to vote only during designated registration periods, only
in person, and only at particular registration centers associated with the polling station and
electoral area for their residence. Someone who wishes to register to vote must declare his or
her name, address, parents’ names, and home area, and the electoral official will fill out this
information on a registration form. The registrant is photographed if a camera is available,
and the photograph is attached to the form, covered by a sticky plastic sleeve and becomes
the official voter ID card.4 Like the United States and several other former British colonies,
Ghana does not have a national ID card system and electoral officials have no means to
check a registrant’s identity, so that it is fairly easy to declare false information. Electoral
officials may remind the person registering that the penalty for giving false information or
registering multiple times is up to a year in prison, but almost no one is ever prosecuted for
false registration.5
Voter registration was delayed several times due to a controversy around a summary of
the 2006 voters register, as well as due to supposed delays in procuring equipment, release of
funds from the government, and hiring of qualified temporary staff.6 Voter registration finally
began on 31 July 2008, with only one day’s advance notice. Although each of the approxi-
mately 4,800 electoral areas in the country was expected to have a registration workstation,
11
only about 2,500 workstations were available. The regional Electoral Commissioners dis-
tributed equipment and registration materials at each regional headquarters to district-level
Electoral Commission officials, who transported them and distributed them at the district
offices to temporary staff hired by the district office. The district-level Electoral Commission
officials then drew up the plans for which electoral areas would have registration centers on
which days. Consequently, no advance information was available centrally on where the mo-
bile registration workstations would be located on which dates. In at least one region, the
distribution of equipment and materials among the numerous districts was haphazard and
did not follow any formula that considered the size of the districts.7 As in previous voter
registration exercises, the political parties actively ferried people to registration centers.8
On the last day of the scheduled 11-day registration period, the Electoral Commission
extended registration by two days due to widespread reports of shortages of materials and
equipment. The Electoral Commission of Ghana then processed all the registration forms at
its headquarters in Accra and produced a provisional voters register. By law, this provisional
register must be made available during an exhibition period, during which an official from
the Electoral Commission sits with the provisional voters register at particular locations
(usually one central location in each electoral area) so that voters can check for their names.
Objections to any names on the register may be lodged with the Electoral Commission at
this time. Approximately 0.4% of new registrations were challenged in 2008, which is ten
times the rate of challenges against new registrations in 2004, and this provisional list was
cut down to a final list of approximately 12.5 million voters. In Ghana, the voters register
is vetted for deceased voters or others who should not be on the register only during this
period between production of the provisional and final voters registers.
The general elections took place as scheduled on 7 December 2008, but no candidate for
president won a majority of votes cast. The presidential run-off election took place as sched-
uled on 28 December 2008 in all areas except opposition NDC-leaning Tain constituency,
12
where there was a shortage of ballot materials. The incumbent NPP initially sought a court
injunction to stop this last election in Tain, but withdrew the challenge and boycotted the
election on 5 January 2009. The opposition NDC won the presidency, with a final official
vote margin of less than 50,000 votes.
Research Design
In consultation with the Coalition of Domestic Election Observers (CODEO), we selected
four of the ten regions of Ghana for this study: Ashanti, Brong Ahafo, Greater Accra,
and Northern Regions. The leadership of CODEO, an umbrella group of 34 major and
many smaller civil society organizations coordinated by the Ghana Center for Democratic
Development (CDD-Ghana), was willing to adjust some of their plans in order to learn about
the effectiveness of their activities. CODEO did not place any restrictions on our choice of
regions, and we selected these four regions in order to cover a wide range of constituencies
within our resource constraints, including several incumbent NPP strongholds, in which the
2004 parliamentary contest was won by the NPP candidate by a 69 point margin; competitive
constituencies; and several opposition NDC strongholds, in which the 2004 parliamentary
contest was won by the NDC candidate by a 50 point margin. Approximately 54% of the
Ghanaian population lived in these four regions as of the last census in 2000, and they
contain 116 of the 230 constituencies and 2,204 of the approximately 4,800 electoral areas
(ELAs, which are subunits of constituencies).
Randomization
Because observers tend to go to locations that are more accessible, are conveniently
located to the last observed location, and are likely to have problems with voter registration,
it is difficult to determine what portion of any observed difference in voter registration
outcomes should be attributed to the presence of observers and what portion to differences
13
in underlying characteristics, even in the absence of spillover effects. We substantially reduce
concerns about confounding by adopting an experimental approach and randomizing which
electoral areas should be observed.
In order to examine spillover effects, we used a two-stage randomized design with blocking
in the first stage in a design similar to Miguel and Kremer (2004). As noted earlier, these
spillovers are forms of interference across units, a violation of the stable unit treatment value
assumption (SUTVA) (Rubin 1978). A simple comparison of means of registration between
treatment and control electoral areas will therefore be a biased estimate of the primary
effect of a registration observer, even if the assignment of observers to registration centers
is randomized. Consequently, we design our experiment to explicitly take into account the
possible strategic response of political parties in a way that allows us to detect both localized
and general spillover effects.
First, within each region, we divided constituencies into blocks according to the difference
in vote share won by the NPP candidate and the NDC candidate in the 2004 parliamentary
elections. Parliamentary constituencies are political units which are not the same as adminis-
trative districts for which government data is made public, and at the time of the experiment,
population and other data were not available at the constituency level. Consequently, we
blocked only on the 2004 elections results in order to improve the efficiency of our estimates.
Within each block, one constituency was randomly assigned to be a treatment constituency
and two others to be control constituencies, so that there are competitive constituencies as
well as stronghold constituencies for each party among both our treatment constituencies
and our control constituencies. Although all regions were available for randomization, as
CODEO’s mission is to organize observers to improve the quality of Ghanaian elections, a
small number of constituencies in some regions were not available for randomization and
exposure to the 2/3 probability that they would not be observed. Those constituencies
14
designated neither treatment nor control were made available for visits by other CODEO
observers not participating in the experiment.
In the second stage, we randomly selected approximately 25% of the electoral areas in
each of the treatment constituencies to be visited by registration observers, which generates
random variation in the number of treatment electoral areas in the neighborhood of an elec-
toral area, conditional on the total number of electoral areas in the neighborhood. Although
we were aware that there would be fewer registration workstations than electoral areas and
that some electoral areas would share workstations, we conducted our randomization over
the list of electoral areas from the 2006 election because the location of the registration
workstations were to be determined by local Electoral Commission officials and unavailable
ahead of time.
Our randomization procedure classifies electoral areas into one of three groups: control
electoral areas in 26 control constituencies, control electoral areas in 13 treatment constituen-
cies, and treatment electoral areas in those 13 treatment constituencies. In the estimation
we take into account this design through the inclusion of the full set of block fixed effects and
correction of standard errors for clustering at the constituency level (Bruhn and McKenzie
2009; Duflo et al. 2007).
Registration observers were recruited from CODEO member organizations that would or-
dinarily field registration observers, were trained together by one of the authors and CODEO
leaders, accredited by the Electoral Commission as regular observers, and deployed at the
same time. The registration observers were instructed to go to their assigned constituencies,
find out where and when the registration centers would be open, and then visit unannounced
only registration workstations in the electoral areas on their list. They were instructed to
visit all the electoral areas on their list at the beginning of the registration period before
revisiting the registration centers once (mostly in rural areas) or twice (mostly in urban
areas) more during the registration period.
15
Registration observers stayed at each registration center for about 1–2 hours on each first
visit and up to a full day in the later visits. In Ghana, observers are not permitted to assist
or interfere with the proceedings, although they may interact with the party agents and
electoral officials at the registration center. The registration observers were asked to fill out
a one-page checklist with questions such as whether the registration center was open upon
the observer’s arrival, whether it had a workstation, whether it was well-marked and easy to
reach, whether there had been any violence, whether the registration center had been forced
to close at any time, whether any people the observer thought were ineligible (e.g., underage)
had been registered, and whether any people the observer thought were eligible had not been
permitted to register. Registration observers were directed to fax back these checklists to
the CODEO secretariat every couple of days or whenever they were in an urban area and
had access to telephones/fax lines.9 CODEO officials read these reports and released one
general press statement during the registration period and one at its conclusion.10
Data
We use a combination of data from our experiment and official sources for our analysis.
We gained access to the official number of registered voters at the polling station-level in 2004
and 2008, and compiled these into electoral area-level figures. We use whether an observer
filed a report for the registration center for a particular electoral area as our measure of
whether that electoral area was visited by a registration observer.
We digitized the Electoral Commission’s map of constituency boundaries. We also
geocoded the 868 electoral areas in our experiment as points by comparing the names of
the polling stations located within those electoral areas with publicly available printed and
digital maps, data from the 2000 population and housing census, coordinates from GPS we
deployed with some of the registration observers, and on occasion, information given by local
electoral officials (Figure 1). We use ArcGIS to calculate the distance between all pairs of
16
electoral areas and construct variables that indicate the number of electoral areas in a 5 km
radius in the same constituency, number of electoral areas in a 10 km radius in the same
constituency, distance to nearest electoral area in the same constituency (km), number of
electoral areas assigned registration observers in a 5 km radius in the same constituency,
number of electoral areas assigned registration observers in a 10 km radius in the same
constituency, and distance to nearest electoral area assigned a registration observer (km).
These variables are summarized in Table 1. We also use these geocoded electoral areas to
make small adjustments to the constituency boundaries so that electoral areas identified
and located by matching polling station names with other sources actually fell within the
boundaries. Neither population figures nor previous elections results were available at this
level of disaggregation.
[Figure 1 about here]
Balance Statistics
We check our randomization procedure with difference in means tests for the baseline
covariates across our three categories of treatment assignment (Table 1). We find that the
density of electoral areas in the neighborhood and the baseline number of registered voters
in the electoral area in 2004 are similar across the three categories of treatment assignment.
We also regress these pre-treatment variables on the constituency-level and electoral area-
level treatment indicators and the full set of block fixed effects (Table 2). We find that the
coefficients on the treatment variables are not statistically significantly different from zero.11
[Tables 1 and 2 about here]
In three of our four regions, the baseline number of registered voters in 2004 do not differ
significantly between treatment and control, nor with electoral areas in constituencies not
selected for the experiment (not shown). However, Trobu-Amasaman constituency in Greater
17
Accra Region, which was assigned to treatment, had approximately 82,000 registered voters
in 17 electoral areas, while the two control constituencies in that block had approximately
110,000 and 140,000 registered voters distributed over 5 and 8 electoral areas, respectively.
Analysis
We examine the effect of registration observers on voter registration using the percentage
increase at the electoral area level between 2004 and 2008 as the outcome, to take account
of the different baseline numbers of registered voters in different locations. The percentage
change in registration from 2004 to 2008 has a mean of 0.257 with a standard deviation of
0.115. Registration data is also available for 2006, but we do not use that data because of
problems with voter registration numbers for 2006 that were found in 2008 and addressed
by the Electoral Commission’s Kangah Commission Report. This commission found that
in several government (NPP) stronghold constituencies the number of registered voters had
doubled between 2004 and 2006. This only became clear after our randomization, and five
of the twelve constituencies in Ashanti Region selected for our experiment were among those
that had nearly double the number of registered voters in 2006 as in 2004. Therefore, we
focus on the change between 2004 and 2008.
Model
To investigate the full model, including the localized spillover effects of observers, we
estimate the following:
Yij = β0+β1Tij+β2Tci +
∑d
(β3d·tdij)+∑d
(β4d·Tijtdij)+∑d
(β5d·ndij)+∑d
(β6d·Tijndij)+µb+εij
(1)
Yij is the percentage change in the number of registered voters from 2004 to 2008 in electoral
area j in constituency i, Tij is an indicator for whether an observer was randomly assigned
18
to electoral area j in constituency i during the registration period in August 2008, T ci is an
indicator for whether an observer was assigned to any of the electoral areas in constituency i
during registration, and µb indicates block fixed effects. Our variable for capturing localized
spillovers is tdij, the number of electoral areas in constituency i assigned to treatment within
distance d of electoral area j in constituency i. ndij is the total number of electoral areas
within distance d of electoral area j in constituency i.
As noted earlier, the stable unit treatment value assumption (SUTVA) is violated in
a world with spillovers, and ignoring these spillovers will result in biased estimates of the
primary effect. Since an electoral area that has neighboring ELAs that are treated is also
always in a treatment constituency, the neighborhood treatment variables (tdij) are positively
correlated with the indicator variable for the treatment constituency (T ci ). Therefore if the
effect of the neighborhood treatment variables (tdij) is non-zero and they are omitted from
the model, the variable T ci is endogenous. If we assume that there is no interaction effect
between T ci and the neighborhood treatment variables tdij so that the coefficients on T c
i ∗ tdij
are zero, then we can determine the direction of the bias resulting from the omission. Since
the correlation between the neighborhood treatment variables tdij and T ci is positive, the sign
will be determined by the true coefficients of the neighborhood treatment variables. Under
our hypothesis of positive localized spillover effects, there will be an upward bias, and if
the true primary effect is negative, a specification that excludes the neighborhood treatment
variables (tdij) will lead to a coefficient that is closer to zero (and potentially insignificant)
because of this upward bias. In the more complex world in which there is feedback among
treatment ELAs, so that the coefficients on T ci ∗tdij are not zero, the neighborhood treatment
variables are still correlated with the treatment ELA indicator T even after the effect of the
constituency-level treatment T ci has been partialled out. Establishing the direction of the
bias in this case is more difficult as it would require knowledge of the covariances of the
19
variables with each other. In either case, excluding the neighborhood variables from the
model specification will lead to biased estimates because of the SUTVA violation.
While our randomization procedure guarantees that any electoral area within a con-
stituency has an equal probability of being selected for treatment, electoral areas do not
all have the same probabilities ex ante of being assigned a particular number of treated
neighbors because some electoral areas are more centrally located than others. In practice,
treatment and control electoral areas within treatment constituencies have the same number
of treated electoral areas in the neighborhood on average (Table 1), but the randomization
procedure does not guarantee that the density of treatment in the neighborhood of an elec-
toral area will be uncorrelated with other characteristics of electoral areas, such as population
density, distance to roads, and other local characteristics, that may affect voter registration.
Therefore, our model also includes ndij, the total number of electoral areas in constituency
i within distance d of electoral area j in constituency i, which will capture all these aspects
that are unrelated to treatment, for which tdij might proxy. As noted earlier, electoral areas
are geographically coded as points, and both t and n are computed as counts of points that
fall within a particular distance d of the given point.
If registration observers deter registration in the electoral areas they visit but these
deterred registrations are displaced to nearby electoral areas, then β1 < 0 and β3d > 0.
We use d = 0005 to denote electoral areas within a 5 km radius and d = 0510 to denote
electoral areas located between 5 km and 10 km from a particular electoral area. We also
add interaction terms between the treatment indicator Tij and tdij and between Tij and ndij
in some specifications.
Table 3 presents our results, with Column 3 reporting results for the main OLS spec-
ification. Allowing the effect of a registration observer in a nearby electoral area to vary
with treatment status and defining “nearby” as a 5 km radius of the electoral area, we
find that within treatment constituencies, electoral areas with registration observers have
20
an approximately 3.5 percentage point smaller increase in registration than electoral areas
without registration observers. At the same time, a registration observer visiting a nearby
electoral area results on average in an approximately 2.7 percentage point greater increase in
registration. This is consistent with the displacement of some of the registration deterred in
the visited electoral areas to nearby electoral areas which experience a greater registration
increase than otherwise. Note that we do not detect the primary effect of observers (T ) or
a general spillover effect (T c) in the model without the neighborhood variables (Column 1).
[Table 3 about here]
In addition to this displacement or localized spillover effect, we detect a general spillover
effect of these registration observers at the constituency level. The coefficients on the
constituency-level and electoral area-level treatment indicators have the same sign, and the
estimate of the former (β2) is 0.041, with a p-value of 0.086 in the OLS specifications. These
results imply that an electoral area assigned a registration observer, but with no electoral
areas assigned registration observers in a 5 km radius (T c = 1, T = 1, t0005 = 0; henceforth
we omit the indices i and j for simplicity), has on average an approximately 7.6 percentage
point smaller increase in registration than an electoral area in a control constituency (T c = 0,
T = 0, t0005 = 0). This average difference shrinks to 5.9 percentage points if an electoral
area in the 5 km neighborhood is assigned a registration observer (T c = 1, T = 1, t0005 = 1)
and to 1.4 percentage points for an electoral area without a registration observer but with
an electoral area in the 5 km neighborhood assigned a registration observer (T c = 1, T = 0,
t0005 = 1). For the specification in Column 3, we reject the null hypothesis of no treatment
effect, H0: β1 = β2 = β3d = β4d = β6d = 0, in a two-tailed test with F (8, 38) = 8.20 and
a p-value of less than 0.0001.12 We also reject the null hypothesis of no primary or general
spillover effect, H0: β1 = β2 = 0, in a two-tailed test with F (2, 38) = 4.82 and a p-value of
0.014.
21
We use our randomized assignment to treatment as an instrument for whether an electoral
area was visited by a registration observer (Vij) in Columns 4–6 of Table 3. Early August
is the end of the rainy season in southern Ghana and still the middle of the rainy season in
northern Ghana, and the registration observers noted difficulty traveling on many rural roads
and crossing rivers. Moreover, there was some confusion surrounding the schedule of which
registration centers would be open on a given date. However, compliance was generally very
good (Table 4) and consequently there are only small differences between our OLS and IV
estimates.
[Table 4 about here]
We check for robustness by estimating equation (1) at other distances (4 and 8 km, 6
and 12 km), including the log number of registered voters in 2004 as a control and using the
log number of registered voters in 2008 as the outcome with the 2004 figures as a control,
and the results remain substantively the same. In general, the estimated primary ITT effect
of registration observers (β1) is about -4% and the estimated localized spillover effect (β3d)
is greater at shorter radii, which is consistent with a displacement of potential registrants
away from an observed registration center to closer alternative registration centers. The
results also do not change substantively when we include an additional “ring” (d = 1020,
for example) to consider the effect of registration observers in electoral areas further away.
We also estimate equation (1) without T c but with constituency fixed effects instead of
block fixed effects, and the results for the primary and spillover effects of observers remain
substantively the same (not shown).
As an additional robustness check, we use the distance to the nearest neighboring treat-
ment electoral area instead of the number of treatment electoral areas in a certain radius.
We regress the percentage change in registration from 2004 to 2008 on the constituency and
electoral area level treatment indicators, the inverse of the distance to the nearest neighbor-
22
ing treatment electoral area ((dTij)−1), its interaction with the constituency-level treatment
indicator variable, and the full set of block fixed effects:
Yij = β0 + β1Tij + β2Tci + β3(d
Tij)−1 + β4T
ci (dTij)
−1 + µb + εij (2)
We expect the effect to be larger the closer the nearest treatment electoral area, since distance
raises the cost of visiting another registration center. We use this transformation of the
distance measure to allow for this effect to diminish more rapidly at closer distances than
at greater distances. We also restrict the sample to electoral areas whose nearest neighbor
treated lies at less than the maximum distance for electoral areas in treatment constituencies
(54 km) for this analysis.
With this set up, the estimated coefficient on the inverse of the distance to the near-
est neighboring treatment electoral area for electoral areas in treatment constituencies is
β3 + β4 = 0.042 with a p-value of 0.028, while the coefficient on distance for electoral areas
in control constituencies (β3 = −0.148) is statistically indistinguishable from zero with a
p-value of 0.490. the estimate for β1 is -0.10 (with a standard error of 0.012) and the co-
efficient on the constituency-level treatment indicator (β2) is estimated to be -0.016 (with
a standard error of 0.018). The estimated coefficient on the constituency-level treatment
indicator (β2) is smaller than our estimates for equation (1) reported in Table 3 because it
effectively incorporates the difference in average distance to the nearest treatment electoral
area between electoral areas in treatment and control constituencies. In another robustness
check, we use the negative log of the distance to the nearest neighboring treatment electoral
area as an alternative measure, and the substantive findings remain the same (not shown).
To get a sense of the magnitude of the treatment effects that includes all localized and
general spillover effects, we perform the following calculation. We set all the treatment vari-
ables (T , T c, and the number of electoral areas assigned registration observers in 0–5 km
23
and 5–10 km (td)) and all the interaction terms involving those variables to zero and then
predict for each electoral area the percentage change in registration using the results from
the baseline regression of Table 3, Column 3. Using this predicted percentage change, we
calculate the predicted absolute change in registrations for the electoral areas in our sample.
These estimates suggest that there would have been 4,600 more registrations in the 13 treat-
ment constituencies in the absence of our treatment. If we assume that the estimates of the
treatment effects apply to the whole country and that these effects remain the same once the
experiment is scaled up (i.e., we treat 25% of electoral areas in each of the 230 constituencies
in Ghana), then we can also quantify the Ghana-wide effect: Because our treatment con-
stituencies contain about 5.3% of all registrants in Ghana in 2004, extrapolating the effect
from the 13 constituencies implies an estimate of 87,000 fewer registrations as a consequence
of a scaled-up treatment. We must be very careful to note, however, that the constituen-
cies in our experiment were not a representative sample of all Ghanaian constituencies and
also that the above calculation assumes, somewhat unrealistically, that increasing the total
number of electoral areas and constituencies treated does not affect the magnitude of the
primary or the spillover effects.
Overall, we find robust evidence that registrations are deterred at treatment electoral
areas and in treatment constituencies more generally. However, some of this reduction in
registration is negated by displacement to electoral areas that are close to treated electoral
areas.
Citizens and Electoral Officials
We revisit two alternative mechanisms by which registration observers might affect regis-
tration, which could complicate our interpretation of the effect of a registration observer on
the electoral area s/he visited as a lower bound estimate on registration irregularities most
likely enabled by party agents.
24
First, registration observers might affect registration by influencing the behavior of citi-
zens who may feel less intimidated and become more likely to register. We believe that this
is unlikely because of the confusion around the schedule and logistics of the registration exer-
cise, as well as the unannounced nature of the registration observers’ visits. Interviews with
district-level Electoral Commission officers indicated that equipment problems sometimes
forced registration centers to be merged and their locations changed with little notice.13
This mechanism cannot account for the positive localized spillover effects presented in Table
3. Moreover, this mechanism would lead us to underestimate the extent to which extra
registrations are deterred by observers, so that the interpretation of our estimate as a lower
bound is still valid.
A second possible mechanism is the influence of registration observers on the behavior of
electoral officials. To investigate this mechanism, we conducted a survey of electoral officials
posted in the electoral areas during the exhibition of the provisional voters register in October
2008. All electoral areas that were selected for treatment during the registration period and
approximately 30% of the remaining electoral areas from both our treatment and control
constituencies were randomly assigned to be visited by observers during the voter register
exhibition period. None of the registration observers were exhibition period observers, so
that no registration observers were involved in evaluating their own effectiveness.
These exhibition period observers conducted a survey of electoral officers and any party
agents present about voter registration in that area and asked for provisional registration
numbers. These observers completed surveys of electoral officers in 304 electoral areas (of
which six are missing information needed to identify the constituency or electoral area). Un-
fortunately, the district-level Electoral Commission officers hire a large number of temporary
staff for these national-scale exercises, and many of the officials posted during the exhibition
period were not the same as those posted during the registration exercise.14
25
Although officially a registrant does not need to present ID in order to be registered,
approximately 60% of exhibition-period electoral officials who reported that there were ob-
servers at the registration center in their electoral area in August asked for ID. A nearly
identical proportion (56%) of these electoral officials who responded that no observers vis-
ited during registration also responded that identification was requested of registrants. Local
electoral officials frequently expounded upon their creative solutions to shortages of registra-
tion forms and malfunctioning registration equipment in order to accommodate the unex-
pectedly large number of people who turned out to register. Registration observers therefore
likely had very little effect on the behavior of electoral officials in ways that depressed reg-
istration. As with the citizens mechanism, it is also difficult to imagine that officiousness
would account for such a large positive localized spillover effect.
Registration observers might alter the behavior of electoral officials in other ways, how-
ever. Electoral officials who see registration observers at a registration center could report
this up the chain of command, affecting the behavior of their counterparts at other regis-
tration centers. This is in accord with the finding of no difference in registration increase
between electoral areas with and without registration observers in the model without the
spillover variables (Column 1) and with the general spillover effect found in the full specifica-
tion (Column 3) in Table 3. However, we have no direct evidence to support this contention,
and by itself, this cannot account for the finding of positive localized spillovers.
The estimated primary effect of registration observers on registration (β1), taking into
account localized spillover effects, may then be interpreted as the effect of registration ob-
servers through their influence on party agents active during registration. We do not argue
that all registration irregularities were deterred where registration observers were present,
and we would underestimate the extent of registration inflation if the registration of eligible
voters was also suppressed. Hence, this is an estimate of the lower bound on the extent of
the registration inflation form of irregularities in Ghana in 2008.
26
Conclusion
This article extends the empirical scholarship on electoral fraud to the study of miscon-
duct at the pre-election stage. It presents findings from a randomized field experiment on
the effect of domestic observers on the extent of irregularities in voter registration in Ghana
in 2008. Our research design and analysis explicitly takes into account the spillovers that
may result from the organization of political parties across multiple electoral areas and the
capacity of party agents to transport supporters from one electoral area to another.
We find a general spillover or constituency-level effect; the increase in the number of regis-
tered voters from 2004 to 2008 was on average 4.1 percentage points smaller for electoral areas
in constituencies with some registration observers than electoral areas in constituencies with
no registration observers. Furthermore, within constituencies with registration observers,
the increase in registration was on average approximately 3.5 percentage points smaller in
electoral areas with observers than without (primary effect). However, an electoral area
with a registration observer located within 5 km led to, on average, a 2.7 percentage point
greater increase in registration. This combination of a positive localized spillover effect from
nearby electoral areas with a negative primary effect is strong evidence that deterred extra
registrations are being displaced. Based upon the design of the experiment, we attribute
this effect to the registration observers’ influence on the activity of party agents. There-
fore, we interpret the negative primary treatment effect as a lower bound on the extent of
irregularities.
This research on irregularities in voter registration has implications for both pro-democracy
actors and scholars of democratization and electoral fraud in partial democracies. As coun-
tries like Ghana are designated “consolidating democracies,” international organizations shift
their attention and resources elsewhere, and the role of domestic observers in protecting
the quality of elections grows in importance.15 Domestic election observers and other pro-
democracy actors may be heartened that a relatively small observer presence had a signifi-
27
cant impact in our experiment. But they should allocate their resources more densely where
greater harmful spillover effects are expected and more towards registration and other ear-
lier stages of the election process (Bjornlund 2004, Carothers 1997). Otherwise, they risk
regularly declaring elections with substantial displaced and hidden fraud as free and fair and
diminishing their credibility and their potential roles in the democratization process.
Furthermore, our finding of positive spillovers has implications for researchers who wish to
measure electoral irregularities or use data collected by observers. Our findings suggest that
even in a model new democracy like Ghana, political parties appear to evade observers, who
deter some but displace substantial irregularities in registration. Although the magnitude of
the primary and spillover effects are likely to vary for other elections with the extent of media
coverage of observers’ activities, the difficulty of reaching an alternative registration center,
and political parties’ resources and past experience, the basic incentive for political parties
to inflate the register and to evade observers while doing so should pertain in elections in
other new democracies. Further investigation of registration and other displaced pre-election
irregularities for additional elections and other countries should improve our measures for
election quality and advance scholarship on electoral fraud and electoral politics in new
democracies.
28
Notes
1We thank the Coalition of Domestic Election Observers and the Center for Democratic
Development in Ghana for their extensive cooperation. We thank Catherine Kelly, Noah
Nathan, Claire Provost, and Jitka Vinduskova for research assistance, and Robert Bates,
Lisa Blaydes, Jorge Domınguez, Adam Glynn, Donald Green, Michael Hiscox, Macartan
Humphreys, Monika Nalepa, Leonard Wantchekon, and Daniel Ziblatt for helpful com-
ments. Earlier versions were presented at the 2009 APSA and ASA meetings, the Uni-
versity of Chicago, Columbia, Harvard, MIT, NYU, and Yale. Support for this research
was provided by the National Science Foundation (SES-0752986), the Weatherhead Center
for International Affairs, the Harvard Academy for International and Area Studies, and the
Milton Fund at Harvard University. Replication data are available at http://www.wiwi.uni-
frankfurt.de/profs/schuendeln/.
2What we label partial democracies are sometimes called hybrid regimes, semi-democracies,
and anocracies, among other terms (Collier and Levitsky 1997; Epstein et al. 2006; Schedler
2002).
3The National Democratic Institute has monitored more than 270 elections (http://www.
ndi.org/content/elections), and the Carter Center and EU have both observed more than
60 elections (http://cartercenter.org/peace/democracy/observed.html; http://ec.europa.eu/
external relations/human rights/election observation/index en.htm).
4The Electoral Commission describes this process as “lamination” but the plastic sleeve
is not heated or melted; it is only self-adhesive.
5A prominent exception is the case of Pius Opoku Boateng, who came under height-
ened scrutiny as the NDC parliamentary candidate for Kwabre West constituency and was
sentenced to 12 months in prison for double registration (Alhassan 2008).
6Meeting with Deputy Electoral Commissioner David Kangah, in Accra, Ghana, July
2008.
7Observation by research assistant at regional Electoral Commission headquarters, 30
July 2008.
8An NDC agent and a taxi driver independently reported to a domestic registration ob-
server in Trobu-Amasaman constituency in Greater Accra Region that, prior to the observer’s
arrival, NPP pick-up trucks conveyed people from nearby villages to the registration center.
29
Similarly in Ningo-Prampram constituency in Greater Accra Region, a domestic registration
observer reported that both NDC and NPP were bussing people to registration centers.
9For previous elections, CODEO trained their registration observers to address these
issues in a 1–2 page written report rather than on a pre-printed checklist with space for
descriptions of any incidents. These observers were free to select which electoral areas to
visit.
10We cannot use CODEO data to investigate effects of the treatment, because these have
no information for the control electoral areas.
11We also formally compare the distribution of the constituency-level vote shares for the
two major parties in the previous presidential election with a Kolmogorov-Smirnov test.
We cannot reject that the distributions of NPP and NDC vote shares are the same for the
whole sample (p = 0.745), among treatment constituencies (p = 0.879), or among control
constituencies (p = 0.918).
12We also use randomization inference (Fisher 1935; Rosenbaum 2002) to test the exact
null of no treatment effect, H0: β1i = β2i = β3di = β4di = β6di = 0, for the intent-to-treat
analysis using radii of 5 and 10 km. Using the F -statistic from the actual experiment and
the randomization procedure from the experiment to generate the null distribution based on
10,000 randomizations, we reject the null in a two-tailed test with an exact p-value of 0.03.
13Telephone interview by research assistant, Greater Accra Region, July 2009.
14Many school teachers and university students were hired for registration since they are
literate and registration took place during the school holidays; they were at school in October.
15Interview with USAID official, Accra, Ghana, July 2008.
30
References
Agence France Press (AFP). 2009. “Yemen Parliament Approves Election De-lay.” February 26, http://www.google.com/hostednews/afp/article/ALeqM5j9jDTS-E f2xYvvwxr45 GErqVAdQ (May 15, 2010).
Alhassan, Issah. 2008. “NDC MP Aspirant Gaoled.” Ghanaian Chronicle, 30 October.
Birch, Sarah. 2007. “Electoral Systems and Election Misconduct.” Comparative PoliticalStudies 40(12):1533–56.
Birch, Sarah. 2008. “Electoral Institutions and Popular Confidence in Electoral Processes:A Cross-National Analysis.” Electoral Studies 27(2):305–20.
Bjornlund, Eric. 2004. Beyond Free and Fair : Monitoring Elections and Building Democ-racy. Baltimore: Johns Hopkins University Press.
Boateng, Michael. 2008. “Two Boys Kidnapped.” Ghanaian Chronicle, August 7.
Bronars, Stephen G. and John R Lott. 1998. “Criminal Deterrence, Geographic Spillovers,and Right-to-Carry Concealed Handgun.” American Economic Review 88(2):475–479.
Bruhn, Miriam and David McKenzie. 2009. “In Pursuit of Balance: Randomization inPractice in Development Field Experiments.” American Economic Journal: AppliedEconomics 1(4):200–232.
Carothers, Thomas. 1997. “The Observers Observed.” Journal of Democracy 8(3):17–31.
Case, William. 1999. “The Philippine Election in 1998: A Question of Quality.” AsianSurvey 39(3):468–485.
Collier, David and Steven Levitsky. 1997. “Democracy with Adjectives: Conceptual Innova-tion in Comparative Research.” World Politics 49(3):430–451.
Collier, Paul and Pedro C. Vicente. 2010. “Votes and Violence: Evidence from a FieldExperiment in Nigeria.” Working paper, Oxford University.
Di Tella, Rafael and Ernesto Schargrodsky. 2004. “Do Police Reduce Crime? EstimatesUsing the Allocation of Police Forces After a Terrorist Attack.” American EconomicReview 94(1):115–33.
Duflo, Esther, Rachel Glennerster and Michael Kremer. 2007. Using Randomization inDevelopment Economics Research: A Toolkit. In Handbook of Development Economics,Vol. 4, ed. T. Paul Schultz and John Strauss. New York: Elsevier pp. 3895–3962.
Eisenstadt, Todd A. 2002. “Measuring Electoral Court Failure in Democratizing Mexico.”International Political Science Review 23(1):47–68.
31
Elklit, Jorgen. 1999. “Electoral Institutional Change and Democratization: You can Lead aHorse to Water, But You Can’t Make it Drink.” Democratization 6(4):28–51.
Elklit, Jorgen and Andrew Reynolds. 2002. “The Impact of Election Administration on theLegitimacy of Emerging Democracies: A New Comparative Politics Research Agenda.”Commonwealth & Comparative Politics 40(2):86–119.
Epstein, David L., Robert Bates, Jack Goldstone, Ida Kristensen and Sharyn O’Halloran.2006. “Democratic Transitions.” American Journal of Political Science 50(3):551–69.
Fisher, Ronald Aylmer. 1935. The Design of Experiments. London: Oliver and Boyd.
Gaige, Fred and John Scholz. 1991. “The 1991 Parliamentary Elections in Nepal: PoliticalFreedom and Stability.” Asian Survey 31(11):1040–1060.
Hartlyn, Jonathan, Jennifer McCoy and Thomas M. Mustillo. 2008. “Electoral GovernanceMatters: Explaining the Quality of Elections in Contemporary Latin America.” Com-parative Political Studies 41(1):73–98.
Humphreys, Macartan, William A. Masters and Martin E. Sandbu. 2006. “The Role ofLeaders in Democratic Deliberation: Results from a Field Experiment in Sao Tome andPrincipe.” World Politics 58(4):583–622.
Hyde, Susan. 2007. “The Observer Effect in International Politics: Evidence from a NaturalExperiment.” World Politics 60(1):37–63.
Kelley, Judith. (Forthcoming). Monitoring Democracy: When International Election Obser-vation Works, and Why it Often Fails. Princeton University Press.
Kitschelt, Herbert and Steven I. Wilkinson, eds. 2007. Patrons, Clients, and Policies: Pat-terns of Democratic Accountability and Political Competition. New York: CambridgeUniversity Press.
Lehoucq, Fabrice. 2003. “Electoral Fraud: Cause, Types, and Consequences.” Annual Reviewof Political Science 6:233–56.
Lehoucq, Fabrice Edouard and Ivan Molina. 2002. Stuffing the Ballot Box: Fraud, ElectoralReform, and Democratization in Costa Rica. New York: Cambridge University Press.
Lindberg, Staffan I. 2006. Democracy and Elections in Africa. Baltimore: Johns HopkinsUniversity Press.
Lindberg, Staffan I. and Minion K. C. Morrison. 2005. “Exploring Voter Alignments inAfrica: Core and Swing Voters in Ghana.” Journal of Modern African Studies 43(4):565– 586.
32
McCann, James A. and Jorge I. Domınguez. 1998. “Mexicans React to Political Fraud andCorruption: An Assessment of Public Opinion and Voting Behavior.” Electoral Studies17(4):483–503.
Mebane, Jr., Walter R. 2006. “Election Forensics: Vote Counts and Benford’s Law.” Pre-sented at the Annual Meeting of the Political Methodology Society.
Miguel, Edward and Michael Kremer. 2004. “Worms: Identifying Impacts on Education andHealth in the Presence of Treatment Externalities.” Econometrica 72(1):159–217.
Myagkov, Mikhail, Peter C. Ordeshook and Dimitri Shakin. 2009. The Forensics of ElectionFraud: Russia and Ukraine. New York: Cambridge University Press.
Nickerson, David W. 2008. “Is Voting Contagious? Evidence from Two Field Experiments.”American Political Science Review 102(1):49–57.
Nugent, Paul. 2001. “Ethnicity as an Explanatory Factor in the Ghana 2000 Elections.”African Issues 29(1/2):2–27.
Olken, Benjamin A. 2010. “Direct Democracy and Local Public Goods: Evidence from aField Experiment in Indonesia.” American Political Science Review 104(2):243–67.
Pastor, R. A. 1999. “The Role of Electoral Administration in Democratic Transitions: Im-plications for Policy and Research.” Democratization 6(4):1–27.
Radio France International (RFI). 2010. “Gbagbo Dissolves Government and Electoral Com-mission.” February 13, http://www.rfi.fr/actuen/articles/122/article 6843.asp (May 17,2010).
Rose, Richard and William Mishler. 2009. “How Do Electors Respond to an “Unfair” Elec-tion? The Experience of Russians.” Post-Soviet Affairs 25(2):118–36.
Rosenbaum, Paul R. 2002. Observational Studies. New York: Springer-Verlag.
Rubin, Donald B. 1978. “Bayesian Inference for Causal Effects: The Role of Randomization.”Annals of Statistics 6:34–48.
Schedler, Andreas. 2002. “The Menu of Manipulation.” Journal of Democracy 13(2):36–50.
Simpser, Alberto. 2010. “More than Winning: The Strategy of Electoral Manipulation.”Manuscript, University of Chicago.
Stokes, Susan C. 2005. “Perverse Accountability: A Formal Model of Machine Politics withEvidence from Argentina.” American Political Science Review 99(3):315–25.
Villalon, Leonardo A. 1994. “Democratizing a (Quasi)Democracy - the Senegalese Electionsof 1993.” African Affairs 93(371):163–193.
33
Wantchekon, Leonard. 2003. “Clientelism and Voting Behavior: Evidence from a FieldExperiment in Benin.” World Politics 55(3):399–422.
Yemen Times. 2009. “The Roots of Protest: Prior Elections Impact Future Polls.” July 5,Document YEMTIM0020090705e57500002, Factiva (May 15, 2010).
Zamble, Fulgence. 2010. “Cote d’Ivoire: Crisis within a Crisis Delays Elections Again.”February 20, http://allafrica.com/stories/201002210001.html (May 17, 2010).
Ziblatt, Daniel. 2006. “How Did Europe Democratize?” World Politics 58(2):311–38.
Ziblatt, Daniel. 2009. “Shaping Democratic Practice and the Causes of Electoral Fraud:Theory and Evidence from Pre-1914 Germany.” American Political Science Review103(1):1–22.
34
Table 1: Means of Variables by Treatment Assignment Status
Assignment Status Difference
T c=1 T c=1 T c=0 (a)–(b) (b)–(c)T=1 (a) T=0 (b) T=0 (c)
Pre-Treatment Variables
# Registered voters in electoral area 1899 2189 1799 -290 390in 2004 (375) (252)
# Electoral areas in 5 km radius in 2.94 3.32 2.79 -0.38 0.53same constituency (0.45) (0.29)
# Electoral areas in 10 km radius in 7.53 7.84 7.22 -0.31 0.62same constituency (0.78) (0.53)
Distance to nearest electoral area in 3.79 4.25 4.31 -0.46 -0.06same constituency (km) (0.82) (0.87)
Spillover Variables
# Electoral areas in 5 km radius 0.75 0.84 0 -0.091 0.844assigned registration observer (0.137) (0.041)
# Electoral areas in 10 km radius 1.95 2.16 0 -0.213 2.16assigned registration observer (0.230) (0.00)
Distance to nearest electoral area 8.34 6.89 41.34 1.45 -34.45assigned a registration observer (km) (1.00) (1.53)
Standard errors in parentheses. N=868 electoral areas.
35
Table 2: Regression of Pre-Treatment Variables on Treatment Assignment
(1) (2) (3)Dependent Variable: # Registered # Electoral # Electoral
Voters in 2004 Areas in 5 km Areas in 10 km
Treatment constituency (T c) -45 0.472 0.619(569) (0.435) (1.119)
Electoral area assigned registration -102 −0.071 0.204observer (T ) (236) (0.377) (0.483)
Block Fixed Effects Yes Yes Yes
R2 0.333 0.335 0.358N 868 868 868
OLS. Disturbances clustered at the constituency level; robust standard errors inparentheses. 39 clusters.
36
Table 3: Effect of Registration Observers on Percentage Change in Registration from 2004to 2008, 5km/10km
Dependent Variable: Percentage change in number of registered voters from 2004 to 2008
(1) (2) (3) (4) (5) (6)OLS OLS OLS IV IV IV
Treatment constituency (T c) −0.006 −0.042+ −0.041+ −0.003 −0.036 −0.036(0.016) (0.022) (0.024) (0.017) (0.022) (0.024)
Electoral area assigned −0.016 −0.030∗ −0.035∗
registration observer (T ) (0.011) (0.014) (0.017)ELA visited by registration −0.022 −0.044∗ −0.042+
observer (V ) (0.015) (0.016) (0.022)# Electoral areas in 5 km 0.028∗ 0.027∗∗
assigned registration observer (0.008) (0.008)# Electoral areas in 5–10 km 0.010 0.011
assigned registration observer (0.006) (0.007)T ∗ # Electoral areas in 5 km −0.007 −0.010
assigned registration observer (0.012) (0.017)T ∗ # Electoral areas in 5–10 km 0.019∗∗∗ 0.013
assigned registration observer (0.004) (0.009)# Electoral areas in 5 km visited 0.023∗ 0.023∗
by registration observer (0.009) (0.010)# Electoral areas in 5-10 km 0.005 0.005
visited by registration observer (0.006) (0.007)V ∗ # Electoral areas in 5 km −0.004 −0.001
visited by registration observer (0.013) (0.019)V ∗ # Electoral areas in 5-10 km 0.019∗∗∗ 0.021
visited by registration observer (0.004) (0.013)# Electoral areas in 5 km 0.001 < 0.001
(0.001) (0.001)# Electoral areas in 5–10 km < −0.001 < −0.001
(0.002) (0.002)T ∗ # Electoral areas in 5 km 0.003
(0.007)T ∗ # Electoral areas in 5–10 km 0.002
(0.005)V ∗ # Electoral areas in 5 km −0.003
(0.009)V ∗ # Electoral areas in 5–10 km < −0.001
(0.008)Block fixed effects Yes Yes Yes Yes Yes Yes
R2 0.166 0.193 0.193 0.159 0.193 0.192N 868 868 868 868 868 868
Disturbances clustered at the constituency level; robust standard errors in parentheses.39 clusters. + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001
37
Table 4: Compliance Rate by Treatment Assignment Status
Assignment Status # Electoral # Visited ComplianceAreas (Vij = 1) (Tij = Vij)
Treatment constituency, treatment electoral area 77 65 84%Treatment constituency, control electoral area 199 24 88%Control constituency, control electoral area 592 1 99%
38
Recommended